Research map / Multi-agent systems and games / Human-AI and goal-conditioned
ERMVP: Communication-Efficient and Collaboration-Robust Multi-Vehicle Perception in Challenging Environments
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ERMVP, a collaborative LiDAR 3D object detection method combining filter-and-merge feature sampling for communication efficiency, consensus-based feature spatial calibration for localization errors, and attention-based fusion with accuracy-enhanced interaction.
Problem
Collaborative multi-vehicle perception suffers from limited communication bandwidth, localization/heading errors causing feature misalignment, and inefficient information aggregation, especially in noisy and bandwidth-limited settings.
Approach
Combines confidence-based foreground filtering with clustering-based merging of feature vectors to reduce both spatial and intra-class redundancy; aligns features via RANSAC-verified matching of sparse consensus regions without pose supervision; and fuses features with alternating local/global attention plus ego-centric accuracy-enhanced cross-attention.
Key results
- On V2V4Real under 0.4/0.4 pose noise, ERMVP reaches 65.01/40.45 AP@0.5/0.7 versus 61.12/32.36 for F-Cooper and 63.69/34.79 for Where2comm (from Table 1 rows).
- With 1.0 m localization error on OPV2V, ERMVP attains 76.09 AP@0.7 versus about 46.66 for the second-best CoAlign (per the text, 29.43% improvement over the second-best method).
- Ablation on V2V4Real (0.4/0.4 noise) shows all three components together raise AP@0.7 from 24.48 (no-collaboration fusion base) to 40.45 with the full model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ERMVP instead of the strongest listed baseline | Collaborative 3D detection · V2V4Real, noise 0.4/0.4 | AP@0.5: 63.33 -> 65.01 | Helped |
| Use ERMVP instead of the strongest listed baseline | Collaborative 3D detection · OPV2V, noise 0.4/0.4 | AP@0.7: 60.64 -> 66.32 | Helped |
| Add attention-based feature fusion (AFF) (ablation) | ERMVP position-wise maximum fusion baseline · V2V4Real, noise 0.4/0.4 | AP@0.5: 51.71 -> 62.24 | Helped |
| Add attention-based feature fusion (AFF) (ablation) | ERMVP position-wise maximum fusion baseline · OPV2V, noise 0.4/0.4 | AP@0.7: 40.12 -> 61.43 | Helped |
| Add feature spatial calibration (FSC) after AFF (ablation) | ERMVP with AFF · V2V4Real, noise 0.4/0.4 | AP@0.5: 62.24 -> 64.36 | Helped |
| Add feature spatial calibration (FSC) after AFF (ablation) | ERMVP with AFF · OPV2V, noise 0.4/0.4 | AP@0.7: 61.43 -> 62.67 | Helped |
| Add accuracy enhanced feature interaction (AEI) after AFF and FSC (ablation) | ERMVP with AFF and FSC · V2V4Real, noise 0.4/0.4 | AP@0.5: 64.36 -> 65.01 | Helped |
| Add accuracy enhanced feature interaction (AEI) after AFF and FSC (ablation) | ERMVP with AFF and FSC · OPV2V, noise 0.4/0.4 | AP@0.7: 62.67 -> 66.32 | Helped |
| Vary the filter rate α and clustering ratio β, which set feature compression (ablation) | ERMVP feature sampling · V2V4Real and OPV2V | AP@0.7 | Hurt |
| Use ERMVP instead of the strongest listed baseline | Collaborative 3D detection · OPV2V, localization noise 1.0 m | AP@0.7: 46.66 -> 76.09 | Helped |
Surprising findings
- Several intermediate fusion methods (e.g., DiscoNet and F-Cooper on OPV2V at 0.4/0.4 noise) perform worse than the No Collaboration baseline under pose noise.
Limitations
- Evaluated only on LiDAR-based 3D detection; other modalities left to future work.
- Performance drops when either filter rate or clustering ratio is reduced under heavy compression.
Experimental setup
Datasets: V2V4Real OPV2V
Benchmarks: V2V4Real OPV2V
Metrics: AP@0.5 AP@0.7 communication volume
Baselines: No Collaboration Late Fusion F-Cooper AttFuse Where2comm DiscoNet V2VNet CoAlign V2X-ViT CoBEVT
Builds on: PointPillar Where2comm V2X-ViT CoBEVT Swin Transformer-style windowed attention
Closest papers by meaning
- RCP-Bench: Benchmarking Robustness for Collaborative Perception Under Diverse Corruptions (CVPR 2025 · similarity 0.88)
- CoSDH: Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization (CVPR 2025 · similarity 0.87)
- STAMP: Scalable Task- And Model-agnostic Collaborative Perception (ICLR 2025 · similarity 0.85)
- AGC-Drive: A Large-Scale Dataset for Real-World Aerial-Ground Collaboration in Driving Scenarios (NeurIPS 2025 · similarity 0.85)
- Privacy-Preserving V2X Collaborative Perception Integrating Unknown Collaborators (AAAI 2025 · similarity 0.85)
- Collaborative Semantic Occupancy Prediction with Hybrid Feature Fusion in Connected Automated Vehicles (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.